· Saravanan Periyasamy, Wind Energy Expert
AI agents raise a wind farm's energy output by giving the O&M team a daily, turbine-by-turbine view of where energy is being lost and why. SanchAI runs a library of agents on wind farms. This article covers four of their jobs:
- Ranking breakdowns by the generation each one puts at risk.
- Separating grid curtailment from turbine derates.
- Flagging turbine underperformance and investigating its root cause.
- Forecasting component temperatures to warn before a derate.
The team decides what to act on. The examples come from one morning report on a 150 MW onshore wind farm of 50 turbines, each rated 3 MW. The farm and its data are illustrative.
Where a wind farm's energy goes
Between the wind that reaches an onshore wind farm and the energy that reaches the meter, 10 to 20% is lost. The layout and the design set wake and electrical losses, and the grid sets curtailment. The O&M team controls two losses. Downtime typically costs 3 to 5% of annual energy. Turbine underperformance costs another 1 to 3% on most farms, and more on some.
A year's energy, from gross potential to the meter
Of the 19.5% lost, 6.0% is controlled by the O&M team: downtime, derates, and turbine underperformance. The rest is set by layout, design, grid, and weather.
How the O&M team works today
A control room, run by the OEM or the owner, watches the farm around the clock, resets faults remotely, and sends technicians to turbines that stay down. Every morning the site manager reviews the daily generation report: yesterday's energy, wind, availability, and the turbines in breakdown. Every month the asset manager reports generation against budget to the owner.
None of these reviews is built to find underperformance. SCADA alarms fire on configured limits, and a turbine 3% below its power curve, a few degrees off the wind, or held at a derate after a grid event trips none of them. Availability still counts it as fully available. Investigating underperformance root causes takes an engineer exporting 10-minute SCADA data and building power curves in Excel, turbine by turbine. So it happens at the monthly review, and only for the turbines whose generation already stands out in the report. The team usually puts whatever the review can't explain down to low wind.
The agents run this analysis every day for every turbine and add the findings to the morning report, next to the breakdown list, so the team starts the day with the root causes already investigated.
Breakdowns ranked by generation at risk
Downtime is the largest loss the O&M team controls, and the team usually works through the breakdown list in the order the faults were logged. The breakdown agent ranks every turbine that is down by the generation it will lose before it is repaired. It estimates that from the forecast wind at the turbine and the repair time of similar faults on the farm. It also ranks the farm's alarm codes by the downtime and MWh each has caused, and marks repeat faults that a remote reset will not fix.
On the wind farm this article follows, four turbines were in breakdown this morning. WTG 23's fault was logged third, but its gearbox oil pressure fault puts the most generation at risk, so it goes to the top. WTG 12's converter has tripped for the third time this month, so it gets a root cause check instead of another reset.
This morning's breakdowns
Curtailment and derates
Yesterday the grid limit on the farm lifted at 14:00. Every turbine returned to full output except WTG 33, which stayed capped at 75% of rated power.
When the grid operator curtails a wind farm, the lost energy counts as curtailment. When a turbine limits its own output, called a derate, the loss belongs to the farm. The cause can be a hot component, a wrong controller parameter, or a power limit left in place after a grid event or a service visit. In SCADA both look the same. The turbine produces less than the wind allows.
The loss attribution agent matches every limited interval against the grid operator's setpoints and the turbine's own status codes, and splits the lost energy into curtailment and derate. It flagged WTG 33 at 14:40, and the control room cleared the limit.
WTG 33 after the grid limit lifted
Power curve underperformance
The turbine underperformance agent compares each turbine with two references. The first is the OEM power curve, corrected to the site's air density. The second is the turbine's working profile, which SanchAI's agentic platform maintains for every WTG from its own best operating period. The OEM curve shows whether a turbine delivers what it was specified to. The working profile shows whether it has changed.
The agent first removes curtailed, derated, and stopped intervals, then builds the curves with the IEC 61400-12-1 binning method.
WTG 08 shows what the agent looks for. Its working profile sits slightly above the OEM curve, because the turbine has always beaten its specification a little. Its 10-minute records now sit up to 10% below that profile between cut-in and rated wind speed, and back on it above rated. The agent reads the shortfall wind speed bin by wind speed bin, which shows where on the curve the energy is being lost.
WTG 08's power curve
- OEM curve, density-corrected
- Working profile
- Actual, binned
- 10-minute records
Placed against both references, the farm's turbines fall into four groups.
Every turbine on the farm
Most of the farm's turbines sit within a percent of both, which is normal scatter. WTG 14 has been below the OEM curve since commissioning and still matches its working profile, a long-standing gap that makes it a candidate for a formal power performance test. WTG 20 still meets the OEM curve but has dropped below its own profile, so it has started to drift. WTG 05, WTG 08, and WTG 11 are below both, which means something has changed on each of them. The agent sizes every shortfall in MWh and revenue and goes on to find the cause.
Anemometer validation
Every expected output depends on the nacelle anemometer. One that reads high makes a healthy turbine look short, and one that reads low hides a real loss. The anemometer analysis agent checks each turbine's wind speed against its neighbours' and against its own power output, and its flags feed into the underperformance agent.
WTG 05's anemometer has read about 3% higher than its neighbours' since it was replaced, while its power matches theirs. Below rated wind speed, power rises with the cube of wind speed, so that 3% makes a healthy turbine look about 5% short. The agent flags the anemometer for a check and takes WTG 05 off the underperformance list.
WTG 05 against its neighbours
Yaw and pitch misalignment
WTG 08 and WTG 11 have healthy anemometers, so the agent looks at yaw misalignment and pitch.
A turbine turns to face the wind using the wind vane on its nacelle. If the vane is misaligned, the turbine runs at an angle to the wind while SCADA shows it facing straight in, and it loses output below rated wind speed. WTG 08 fits that pattern, as its power curve showed. The agent groups its records by vane reading and finds that its power peaks when the vane reads about 15°, not 0°. An offset that large is unusual, and it costs about 10% of the turbine's power below rated. The vane needs recalibrating.
WTG 08, power by wind vane reading
WTG 11's shortfall started the day after a pitch system service visit and sits between 7 and 11 m/s. The agent reads the pitch angle of each blade and finds all three holding about 1.5° above the minimum pitch angle they ran at before the visit, so the rotor captures less energy below rated. The cause is in the pitch setup from that visit: a calibration offset or a wrong parameter set. The agent also compares the three blades with each other on every turbine. A single blade out of line costs little energy but loads the rotor unevenly, and the pitch motor current and the blade load sensors show it. That finding goes to the pitch team as a loads issue.
Step changes like these, after a sensor replacement, a service visit, or a controller update, show up in the data within days. Slow losses such as blade leading-edge erosion still take months of data to separate from normal scatter, and the agent tracks those against the working profile over the season.
Component temperature forecasting
The component temperature agent compares several critical drivetrain and electrical components with the temperature each normally runs at for the same load and ambient temperature. It forecasts the next 3 days of temperatures for each component, using the weather forecast, and flags any component expected to reach its derate threshold, the temperature at which the turbine cuts its own output to protect the component.
WTG 17's gearbox bearing has run hotter than normal for several days, and the gap is growing. Stronger wind is forecast in two days, and the agent expects the bearing to cross the threshold then. It asks for a check of the oil cooler and lubrication before the high-wind spell, when lost output would cost the most.
WTG 17's gearbox bearing
- WTG 17, measured
- WTG 17, forecast
- Healthy bearing at the same load
The morning report
The report lists every finding above as one item, with the anemometer flags already applied. Each item carries the probable cause, the recommended check, and the generation at risk in MWh. The site manager approves or rejects each item, and later reports show whether each shortfall has closed.
This morning's report
Morning report 150 MW, 50 turbines · 06:30
Approved Pending
Breakdowns MWh until repair
Underperformance MWh per month
Component temperature MWh at risk
What changes for the team
Across a farm, catching yaw, pitch, and derate problems as they appear is worth 0.5 to 2% of annual generation, depending on the farm and its starting point. The report prices every item in MWh and revenue, so the asset manager sees what a fix is worth before approving it.
Before and after the agents
Frequently asked questions
Can the agents prove that a turbine is below its warranted power curve?
No. Nacelle anemometer data is enough to find a shortfall, not to prove a warranty claim. That needs a formal power performance test with a met mast or lidar. The report shows which turbines justify one.
How do the agents tell a low-wind month from a turbine problem?
At month end, the wind resource agent compares the month's wind with a 20-year ERA5 baseline and splits the gap to budget into wind, downtime, curtailment, and performance. A month that missed budget because the wind was below its long-term average is explained by the wind. A month that missed on normal wind has its cause in one of the other buckets, and the daily reports already show which one.
Do the agents replace the control room?
No. The control room keeps turbines running and resets faults. The agents work on the same SCADA data to find the energy losses the control room is not set up to look for, and report them to the O&M team. They read SCADA and never write to it.
What data do the agents need?
The 10-minute SCADA data the turbines already record, with the alarm and event logs. Maintenance records and work orders add context when they are available. No new sensors are needed.
How long does a pilot take?
A pilot takes 2 weeks on your wind farm's data, and full deployment takes about 4 weeks.
Can we set up our own agents?
Yes. Describe the job once in plain language, for example 'Report the top 5 performing turbines every day at 8:00 AM for the previous 24 hours,' and SanchAI runs it on that schedule.
Saravanan Periyasamy is SanchAI's wind domain expert. He has more than 24 years of experience in the wind industry, including at Vestas and Senvion.
Find the turbines losing energy on your wind farm
SanchAI runs these agents every day across a 2 GW portfolio of wind, solar, and BESS plants.
Start a Pilot